# Title, CIDR ranges, arrow labels and flow steps for the gallery diagram. Use with: # terravision draw --source aws-sagemaker-mlops.tvg.json --annotate aws-sagemaker-mlops.annotations.yml format: '0.3' title: Amazon SageMaker MLOps with a private Studio, Pipelines and a real-time endpoint update: aws_vpc.ml: cidr_block: 10.30.0.0/16 aws_subnet.studio~1: cidr_block: 10.30.16.0/20 aws_subnet.inference~1: cidr_block: 10.30.64.0/24 aws_subnet.endpoints~1: cidr_block: 10.30.240.0/24 aws_subnet.studio~2: cidr_block: 10.30.32.0/20 aws_subnet.inference~2: cidr_block: 10.30.65.0/24 aws_subnet.endpoints~2: cidr_block: 10.30.241.0/24 aws_sagemaker_domain.studio~1: label: SageMaker Studio aws_sagemaker_domain.studio~2: label: SageMaker Studio aws_sagemaker_pipeline.churn_training: label: SageMaker Pipelines Churn Training aws_sagemaker_feature_group.customer_features: label: Feature Store Customer Features aws_sagemaker_model_package_group.churn_models: label: Model Registry Churn Models connect: aws_ec2_client_vpn_endpoint.data_science: - aws_sagemaker_domain.studio~1: VPC-only Studio aws_scheduler_schedule.weekly_retrain: - aws_sfn_state_machine.retrain: Weekly aws_sfn_state_machine.retrain: - aws_sagemaker_pipeline.churn_training: StartPipelineExecution - aws_sagemaker_endpoint.churn~1: Blue/green deploy - aws_sagemaker_endpoint.churn~2: Blue/green deploy aws_sagemaker_pipeline.churn_training: - aws_sagemaker_model_package_group.churn_models: Register version aws_vpc_endpoint.sagemaker_runtime~1: - aws_sagemaker_endpoint.churn~1: InvokeEndpoint aws_vpc_endpoint.sagemaker_runtime~2: - aws_sagemaker_endpoint.churn~2: InvokeEndpoint flows: Experiment: description: A data scientist builds features steps: - resource: tv_aws_users.data_scientists -> aws_ec2_client_vpn_endpoint.data_science xlabel: VPN detail: Data scientists connect through Client VPN - resource: aws_ec2_client_vpn_endpoint.data_science -> aws_sagemaker_domain.studio~1 xlabel: Studio detail: to SageMaker Studio, which runs in VPC-only mode with no internet access - resource: aws_sagemaker_domain.studio~1 -> aws_vpc_endpoint.sagemaker_api~1 xlabel: API detail: Notebooks call SageMaker through the interface endpoint - resource: aws_vpc_endpoint.sagemaker_api~1 -> aws_sagemaker_feature_group.customer_features xlabel: Features detail: to define and fill feature groups in Feature Store Retrain and deploy: description: A new model reaches production steps: - resource: aws_scheduler_schedule.weekly_retrain -> aws_sfn_state_machine.retrain xlabel: Start detail: An EventBridge Scheduler schedule starts the Step Functions workflow each week - resource: aws_sfn_state_machine.retrain -> aws_sagemaker_pipeline.churn_training xlabel: Train detail: which runs the SageMaker pipeline - resource: aws_sagemaker_pipeline.churn_training -> aws_sagemaker_feature_group.customer_features xlabel: Read detail: The pipeline trains on offline features and the training data in S3 - resource: aws_sagemaker_pipeline.churn_training -> aws_sagemaker_model_package_group.churn_models xlabel: Register detail: and registers the evaluated model in the Model Registry - resource: aws_sfn_state_machine.retrain -> aws_sagemaker_endpoint.churn~1 xlabel: Deploy detail: Once the version is approved, the workflow updates the endpoint with a blue/green deployment Prediction: description: The app asks for a churn score steps: - resource: tv_aws_mobile_client.customer_app -> aws_api_gateway_rest_api.predictions xlabel: Request detail: The customer app calls the predictions API - resource: aws_api_gateway_rest_api.predictions -> aws_lambda_function.predict~1 xlabel: Invoke detail: API Gateway invokes a Lambda function in the VPC - resource: aws_lambda_function.predict~1 -> aws_vpc_endpoint.sagemaker_runtime~1 xlabel: Score detail: which calls the SageMaker runtime through its interface endpoint - resource: aws_vpc_endpoint.sagemaker_runtime~1 -> aws_sagemaker_endpoint.churn~1 xlabel: Infer detail: and the real-time endpoint returns the score